OpenACC: Allow implicit uses of assumed-size arrays in offload regions

This patch reimplements the functionality of the previously-reverted
patch "Assumed-size arrays with non-lexical data mappings". The purpose
is to support implicit uses of assumed-size arrays for Fortran when those
arrays have already been mapped on the target some other way (e.g. by
"acc enter data").

This relates to upstream OpenACC issue 489 (not yet resolved).

gcc/fortran/
	* trans-openmp.cc (gfc_omp_finish_clause): Treat implicitly-mapped
	assumed-size arrays as zero-sized for OpenACC, rather than an error.

gcc/testsuite/
	* gfortran.dg/goacc/assumed-size.f90: Don't expect error.

libgomp/
	* testsuite/libgomp.oacc-fortran/nonlexical-assumed-size-1.f90: New
	test.
	* testsuite/libgomp.oacc-fortran/nonlexical-assumed-size-2.f90: New
	test.

4 files changed, 84 insertions, 6 deletions

diff --git a/gcc/fortran/trans-openmp.cc b/gcc/fortran/trans-openmp.cc
index c28d937..cfecdb0 100644
--- a/gcc/fortran/trans-openmp.cc
+++ b/gcc/fortran/trans-openmp.cc
@@ -1595,6 +1595,7 @@ gfc_omp_finish_clause (tree c, gimple_seq *pre_p, bool openacc)
 
   tree decl = OMP_CLAUSE_DECL (c);
   location_t loc = OMP_CLAUSE_LOCATION (c);
+  bool assumed_size = false;
 
   /* Assumed-size arrays can't be mapped implicitly, they have to be
      mapped explicitly using array sections.  */
@@ -1605,9 +1606,14 @@ gfc_omp_finish_clause (tree c, gimple_seq *pre_p, bool openacc)
 				GFC_TYPE_ARRAY_RANK (TREE_TYPE (decl)) - 1)
 	 == NULL)
     {
-      error_at (OMP_CLAUSE_LOCATION (c),
-		"implicit mapping of assumed size array %qD", decl);
-      return;
+      if (openacc)
+	assumed_size = true;
+      else
+	{
+	  error_at (OMP_CLAUSE_LOCATION (c),
+		    "implicit mapping of assumed size array %qD", decl);
+	  return;
+	}
     }
 
   if (OMP_CLAUSE_MAP_KIND (c) == GOMP_MAP_FORCE_DEVICEPTR)
@@ -1662,7 +1668,9 @@ gfc_omp_finish_clause (tree c, gimple_seq *pre_p, bool openacc)
       else
 	{
 	  OMP_CLAUSE_DECL (c) = decl;
-	  OMP_CLAUSE_SIZE (c) = NULL_TREE;
+	  OMP_CLAUSE_SIZE (c) = assumed_size ? size_zero_node : NULL_TREE;
+	  if (assumed_size)
+	    OMP_CLAUSE_MAP_MAYBE_ZERO_LENGTH_ARRAY_SECTION (c) = 1;
 	}
       if (TREE_CODE (TREE_TYPE (orig_decl)) == REFERENCE_TYPE
 	  && (GFC_DECL_GET_SCALAR_POINTER (orig_decl)
diff --git a/gcc/testsuite/gfortran.dg/goacc/assumed-size.f90 b/gcc/testsuite/gfortran.dg/goacc/assumed-size.f90
index 4fced2e..12f44c4 100644
--- a/gcc/testsuite/gfortran.dg/goacc/assumed-size.f90
+++ b/gcc/testsuite/gfortran.dg/goacc/assumed-size.f90
@@ -4,7 +4,8 @@
 ! exit data, respectively.
 
 ! This does not appear to be supported by the OpenACC standard as of version
-! 3.0.  Check for an appropriate error message.
+! 3.0.  There is however real-world code that relies on this working, so we
+! make an attempt to support it.
 
 program test
   implicit none
@@ -26,7 +27,6 @@ subroutine dtest (a, n)
   !$acc enter data copyin(a(1:n))
 
   !$acc parallel loop
-! { dg-error {implicit mapping of assumed size array 'a'} "" { target *-*-* } .-1 }
   do i = 1, n
      a(i) = i
   end do
diff --git a/libgomp/testsuite/libgomp.oacc-fortran/nonlexical-assumed-size-1.f90 b/libgomp/testsuite/libgomp.oacc-fortran/nonlexical-assumed-size-1.f90
new file mode 100644
index 0000000..8b173c7
--- /dev/null
+++ b/libgomp/testsuite/libgomp.oacc-fortran/nonlexical-assumed-size-1.f90
@@ -0,0 +1,29 @@
+! { dg-do run }
+
+program p
+implicit none
+integer :: myarr(10)
+
+myarr = 0
+
+call subr(myarr)
+
+if (myarr(5).ne.5) stop 1
+
+contains
+
+subroutine subr(arr)
+implicit none
+integer :: arr(*)
+
+!$acc enter data copyin(arr(1:10))
+
+!$acc serial
+! { dg-warning {using .vector_length \(32\)., ignoring 1} "" { target openacc_nvidia_accel_selected } .-1 }
+arr(5) = 5
+!$acc end serial
+
+!$acc exit data copyout(arr(1:10))
+
+end subroutine subr
+end program p
diff --git a/libgomp/testsuite/libgomp.oacc-fortran/nonlexical-assumed-size-2.f90 b/libgomp/testsuite/libgomp.oacc-fortran/nonlexical-assumed-size-2.f90
new file mode 100644
index 0000000..659fe8e
--- /dev/null
+++ b/libgomp/testsuite/libgomp.oacc-fortran/nonlexical-assumed-size-2.f90
@@ -0,0 +1,41 @@
+! { dg-do run }
+
+program p
+implicit none
+integer :: myarr(10)
+
+myarr = 0
+
+call subr(myarr)
+
+if (myarr(5).ne.5) stop 1
+
+contains
+
+subroutine subr(arr)
+implicit none
+integer :: arr(*)
+
+! At first glance, it might not be obvious how this works.  The "enter data"
+! and "exit data" operations expand to a pair of mapping nodes for OpenACC,
+! GOMP_MAP_{TO/FROM} and GOMP_MAP_POINTER.  The former maps the array data,
+! and the latter creates a separate mapping on the target for the pointer
+! itself with a bias so it represents the "zeroth" element.
+
+!$acc enter data copyin(arr(2:8))
+
+! ...then this implicit mapping creates a zero-length array section
+! (GOMP_MAP_ZERO_LEN_ARRAY_SECTION) followed by another GOMP_MAP_POINTER for
+! 'arr'.  But now that pointer is already "present" on the target, so is not
+! overwritten.
+
+!$acc serial
+! { dg-warning {using .vector_length \(32\)., ignoring 1} "" { target openacc_nvidia_accel_selected } .-1 }
+! This access is then done via the on-target pointer.
+arr(5) = 5
+!$acc end serial
+
+!$acc exit data copyout(arr(2:8))
+
+end subroutine subr
+end program p